huskies: merge 954

server/src/agents/pool/pipeline/advance/mod.rs

@@ -74,7 +74,58 @@ impl AgentPool {
                 // Supervisors and unknown agents do not advance the pipeline.
             }
             PipelineStage::Coder => {
-                if completion.gates_passed {
+                if completion.needs_commit_recovery {
+                    // The coder exited with uncommitted content but no commits.
+                    // Check if this is already a second recovery attempt (the
+                    // first recovery respawn also produced no commits).
+                    let recovery_key = format!("{story_id}:commit_recovery_pending");
+                    if crate::db::read_content(&recovery_key).is_some() {
+                        // Second attempt still produced no commits → block.
+                        crate::db::delete_content(&recovery_key);
+                        slog!(
+                            "[pipeline] Coder '{agent_name}' (commit-recovery respawn) \
+                             still produced no commits for '{story_id}'. Blocking story."
+                        );
+                        let reason = "agent declined to commit recoverable work".to_string();
+                        if let Err(e) =
+                            crate::agents::lifecycle::transition_to_blocked(story_id, &reason)
+                        {
+                            slog_error!("[pipeline] Failed to block '{story_id}': {e}");
+                        }
+                        let _ = self.watcher_tx.send(WatcherEvent::StoryBlocked {
+                            story_id: story_id.to_string(),
+                            reason,
+                        });
+                    } else {
+                        // First occurrence: issue a commit-only recovery respawn.
+                        // This does NOT consume a retry_count slot.
+                        crate::db::write_content(&recovery_key, "1");
+                        slog!(
+                            "[pipeline] Coder '{agent_name}' exited with uncommitted work \
+                             for '{story_id}'. Issuing commit-only recovery respawn."
+                        );
+                        let addendum = "\n\nYou have uncommitted work in this worktree. \
+                            Your only task this session is run_tests → git_add → git_commit. \
+                            Do not explore further.";
+                        if let Err(e) = self
+                            .start_agent(
+                                &project_root,
+                                story_id,
+                                Some(agent_name),
+                                Some(addendum),
+                                previous_session_id,
+                            )
+                            .await
+                        {
+                            slog_error!(
+                                "[pipeline] Failed to start commit-recovery respawn \
+                                 for '{story_id}': {e}"
+                            );
+                        }
+                    }
+                } else if completion.gates_passed {
+                    // Clear any stale recovery key when the coder succeeds normally.
+                    crate::db::delete_content(&format!("{story_id}:commit_recovery_pending"));
                     // Determine effective QA mode for this story.
                     let qa_mode = {
                         let item_type = crate::agents::lifecycle::item_type_from_id(story_id);
@@ -130,6 +181,9 @@ impl AgentPool {
                         }
                     }
                 } else {
+                    // Clear any stale recovery key when gates fail normally (agent committed
+                    // but the build is broken — treat as a standard retry, not a recovery).
+                    crate::db::delete_content(&format!("{story_id}:commit_recovery_pending"));
                     // Bug 645 / 668: Before retry/block, check if the agent left committed
                     // work AND the agent had a passing run_tests result captured during its
                     // session.  An agent may crash mid-output (e.g. Claude Code CLI PTY write